Language Model Dictionaries for Text Predictions

Abstract

Techniques are described to generate text prediction candidates corresponding to detected text characters according to an adaptive language model that includes multiple individual language model dictionaries. Respective scoring data from the dictionaries is combined to select prediction candidates in different interaction scenarios. In an implementation, dictionaries corresponding to multiple different languages are combined to produce multi-lingual predictions. Predictions for different languages may be weighted proportionally according to relative usage by a user. Weights used to combine contributions from multiple dictionaries may also depend upon factors such as how recently a word is used, number of times used, and so forth. Further, the dictionaries may include interaction-specific dictionaries that are learned by monitoring a user's typing activity to adapt predictions to corresponding usage scenarios. Interaction-specific dictionaries may be applied selectively for predictions in respective usage scenarios, including interaction with a particular application, application type, person, contact group, or location.

1 . A method comprising:
 recognizing use of multiple different languages for text input in a particular interaction scenario;   activating multiple dictionaries corresponding to the multiple different languages to employ for text predictions in connection with the interaction scenario; and   generating multi-lingual text predictions for text entry associated with the interaction scenario by combining word probabilities obtained using the multiple dictionaries according to an adaptive language model.   
     
     
         2 . A method as recited in  claim 1 , wherein activating multiple dictionaries comprises locating and installing a general population dictionary for a secondary language along-side an existing general population dictionary for the adaptive language model. 
     
     
         3 . A method as recited in  claim 1 , wherein activating multiple dictionaries comprises creating a language specific dictionary for a secondary language as part of the adaptive language model. 
     
     
         4 . A method as recited in  claim 1 , wherein combining word probabilities obtained using the multiple dictionaries comprises interpolating individual word probabilities from the multiple dictionaries using interpolation weights for contributions of each of the multiple dictionaries that are computed according to an interpolation function. 
     
     
         5 . A method as recited in  claim 4 , wherein the interpolation weights selected for the interpolation are proportional to relative usage of the multiple different languages for the particular interaction scenario. 
     
     
         6 . A method as recited in  claim 4 , wherein the interpolation function is configured to vary the weights to increase the contribution of words that are used more recently. 
     
     
         7 . A method as recited in  claim 1 , wherein generating the multi-lingual text predictions comprises:
 generating language specific lists of prediction candidates separately for each of the multiple dictionaries; and   interpolating individual probabilities from the language specific lists to form a common list of the multi-lingual text predictions.   
     
     
         8 . A method comprising:
 identifying an interaction-specific dictionary associated with text input for an interaction scenario with a computing device; and   computing one or more text predictions for the interaction scenario using word probabilities from the interaction-specific dictionary as a component of probabilities assigned by an adaptive language model to determine the one or more text predictions.   
     
     
         9 . A method as recited in  claim 8 , wherein:
 the interaction scenario is defined according to one or more usage parameters indicative of characteristics of user interaction with the computing device; and   identifying the interaction-specific dictionary includes recognizing a current interaction as matching the interaction scenario based upon usage parameters established for the current interaction.   
     
     
         10 . A method as recited in  claim 9 , wherein the usage parameters that define the interaction scenario comprise one or more of an application identity, a type of application, a person, a time of day, a date, a geographic location, a time of year, a setting, a topic, or a particular language used for the interaction scenario. 
     
     
         11 . A method as recited in  claim 8 , further comprising:
 collecting data indicative of a user's typing style for the interaction scenario; and   creating the interaction-specific dictionary for the interaction scenario to contain scoring data for words input in connection with the interaction scenario using the collected data indicative of the user's typing style.   
     
     
         12 . A method as recited in  claim 8 , wherein the interaction-specific dictionary is included in a hierarchy of language model dictionaries established by the adaptive language model to adapt text predictions to a user's individual style in different scenarios. 
     
     
         13 . A method as recited in  claim 8 , wherein computing the one or more text predictions comprises:
 computing scores for candidate words text predictions as weighted combinations of the word probabilities from the interaction-specific dictionary and word probabilities from a general population dictionary representative of common usage across a community of users; and   ranking the candidate words one to another based on the computed scores.   
     
     
         14 . A method as recited in  claim 8 , wherein the interaction-specific dictionary corresponds to interaction via the computing device with at least one of a particular person, application, or location. 
     
     
         15 . A method as recited in  claim 8 , wherein the interaction scenario corresponds to interaction via the computing device with a group of applications having a same application type. 
     
     
         16 . A method as recited in  claim 8 , wherein the interaction scenario corresponds to interaction via the computing device with a contact group associated with a user's contacts. 
     
     
         17 . A computing device comprising:
 a processing system; and   one or more computer-readable media storing instructions that, when executed by the processing system, implement a text prediction engine operable to:
 collect data indicative of a user's typing style for a particular interaction scenario defined according to usage parameters indicative of characteristics of interaction with the computing device; 
 create an interaction-specific dictionary for the particular interaction scenario containing conditional probabilities for words input by the user using the collected data indicative of the user's typing style; 
 detect text input for a subsequent interaction that matches the particular interaction scenario; and 
 interpolate conditional probabilities corresponding to text characters input during the subsequent interaction from the interaction-specific dictionary and at least one other dictionary available to the text prediction engine to generate one or more predictions for the input text characters. 
   
     
     
         18 . A computing device as recited in  claim 17 , wherein the interaction-specific dictionary corresponds to interaction via the computing device with at least one of a particular person, application, or location. 
     
     
         19 . A computing device as recited in  claim 17 , wherein the interaction-specific dictionary corresponds to interaction via the computing device with a particular person and a particular application. 
     
     
         20 . A computing device as recited in  claim 17 , wherein the other dictionary available to the text prediction engine comprises a general population dictionary representative of common usage across a community of users.